Two-View Mammographic Mass Retrieval Using Deep Hashing

Breast cancer is the most commonly diagnosed cancer among women all over the world. Content-based mammographic mass retrieval can assist radiologists to retrieve biopsy-proven masses content similar to the diagnostic ones, which can help radiologists to improve the diagnostic performance. However, although screening mammograms usually consists of two-view (mediolateral oblique view and craniocaudal view) mammography of the same breast, most breast computer-aided diagnosis(CAD) methods incorporated with content-based image retrieval (CBIR) techniques are based on a single-view principle where query mass within a view is analyzed independently. With the advance of deep neural networks in image representation, the deep hashing technique for CBIR maps images into binary feature vectors, which can greatly reduce time and memory costs. In this paper, according to our previous studies, we propose a two-view mammographic mass retrieval approach based on deep hashing. In our retrieval paradigm, a convolutional neural network(CNN) is used to learn both image representation and hash function based on paired labels and class labels. The last layer of CNN will output the image binary codes. Subsequently, the learned binary codes of mass pair are used for matching and ranking within a traditional query-by-example(QBE) framework. Using a dataset that contains a total of 3000 masses collected from the public Digital Database for Screening Mammography(DDSM), this method can yield a retrieval accuracy of 0.8975, which demonstrates its encouraging performance.